1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
52 #include <asm/uaccess.h>
54 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
55 #define MEM_CGROUP_RECLAIM_RETRIES 5
56 struct mem_cgroup *root_mem_cgroup __read_mostly;
58 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
59 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
60 int do_swap_account __read_mostly;
61 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
63 #define do_swap_account (0)
67 * Per memcg event counter is incremented at every pagein/pageout. This counter
68 * is used for trigger some periodic events. This is straightforward and better
69 * than using jiffies etc. to handle periodic memcg event.
71 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
73 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
74 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
77 * Statistics for memory cgroup.
79 enum mem_cgroup_stat_index {
81 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
83 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
84 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
85 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
86 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */
87 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */
91 MEM_CGROUP_STAT_NSTATS,
94 struct mem_cgroup_stat_cpu {
95 s64 count[MEM_CGROUP_STAT_NSTATS];
99 * per-zone information in memory controller.
101 struct mem_cgroup_per_zone {
103 * spin_lock to protect the per cgroup LRU
105 struct list_head lists[NR_LRU_LISTS];
106 unsigned long count[NR_LRU_LISTS];
108 struct zone_reclaim_stat reclaim_stat;
109 struct rb_node tree_node; /* RB tree node */
110 unsigned long long usage_in_excess;/* Set to the value by which */
111 /* the soft limit is exceeded*/
113 struct mem_cgroup *mem; /* Back pointer, we cannot */
114 /* use container_of */
116 /* Macro for accessing counter */
117 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
119 struct mem_cgroup_per_node {
120 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
123 struct mem_cgroup_lru_info {
124 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
128 * Cgroups above their limits are maintained in a RB-Tree, independent of
129 * their hierarchy representation
132 struct mem_cgroup_tree_per_zone {
133 struct rb_root rb_root;
137 struct mem_cgroup_tree_per_node {
138 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
141 struct mem_cgroup_tree {
142 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
145 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
147 struct mem_cgroup_threshold {
148 struct eventfd_ctx *eventfd;
153 struct mem_cgroup_threshold_ary {
154 /* An array index points to threshold just below usage. */
155 int current_threshold;
156 /* Size of entries[] */
158 /* Array of thresholds */
159 struct mem_cgroup_threshold entries[0];
162 struct mem_cgroup_eventfd_list {
163 struct list_head list;
164 struct eventfd_ctx *eventfd;
167 static void mem_cgroup_threshold(struct mem_cgroup *mem);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
171 * The memory controller data structure. The memory controller controls both
172 * page cache and RSS per cgroup. We would eventually like to provide
173 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
174 * to help the administrator determine what knobs to tune.
176 * TODO: Add a water mark for the memory controller. Reclaim will begin when
177 * we hit the water mark. May be even add a low water mark, such that
178 * no reclaim occurs from a cgroup at it's low water mark, this is
179 * a feature that will be implemented much later in the future.
182 struct cgroup_subsys_state css;
184 * the counter to account for memory usage
186 struct res_counter res;
188 * the counter to account for mem+swap usage.
190 struct res_counter memsw;
192 * Per cgroup active and inactive list, similar to the
193 * per zone LRU lists.
195 struct mem_cgroup_lru_info info;
198 protect against reclaim related member.
200 spinlock_t reclaim_param_lock;
202 int prev_priority; /* for recording reclaim priority */
205 * While reclaiming in a hierarchy, we cache the last child we
208 int last_scanned_child;
210 * Should the accounting and control be hierarchical, per subtree?
216 unsigned int swappiness;
217 /* OOM-Killer disable */
218 int oom_kill_disable;
220 /* set when res.limit == memsw.limit */
221 bool memsw_is_minimum;
223 /* protect arrays of thresholds */
224 struct mutex thresholds_lock;
226 /* thresholds for memory usage. RCU-protected */
227 struct mem_cgroup_threshold_ary *thresholds;
230 * Preallocated buffer to be used in mem_cgroup_unregister_event()
231 * to make it "never fail".
232 * It must be able to store at least thresholds->size - 1 entries.
234 struct mem_cgroup_threshold_ary *__thresholds;
236 /* thresholds for mem+swap usage. RCU-protected */
237 struct mem_cgroup_threshold_ary *memsw_thresholds;
239 /* the same as __thresholds, but for memsw_thresholds */
240 struct mem_cgroup_threshold_ary *__memsw_thresholds;
242 /* For oom notifier event fd */
243 struct list_head oom_notify;
246 * Should we move charges of a task when a task is moved into this
247 * mem_cgroup ? And what type of charges should we move ?
249 unsigned long move_charge_at_immigrate;
253 struct mem_cgroup_stat_cpu *stat;
256 /* Stuffs for move charges at task migration. */
258 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
259 * left-shifted bitmap of these types.
262 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
263 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
267 /* "mc" and its members are protected by cgroup_mutex */
268 static struct move_charge_struct {
269 struct mem_cgroup *from;
270 struct mem_cgroup *to;
271 unsigned long precharge;
272 unsigned long moved_charge;
273 unsigned long moved_swap;
274 struct task_struct *moving_task; /* a task moving charges */
275 wait_queue_head_t waitq; /* a waitq for other context */
277 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
280 static bool move_anon(void)
282 return test_bit(MOVE_CHARGE_TYPE_ANON,
283 &mc.to->move_charge_at_immigrate);
286 static bool move_file(void)
288 return test_bit(MOVE_CHARGE_TYPE_FILE,
289 &mc.to->move_charge_at_immigrate);
293 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
294 * limit reclaim to prevent infinite loops, if they ever occur.
296 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
297 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
300 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
301 MEM_CGROUP_CHARGE_TYPE_MAPPED,
302 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
303 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
304 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
305 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
309 /* only for here (for easy reading.) */
310 #define PCGF_CACHE (1UL << PCG_CACHE)
311 #define PCGF_USED (1UL << PCG_USED)
312 #define PCGF_LOCK (1UL << PCG_LOCK)
313 /* Not used, but added here for completeness */
314 #define PCGF_ACCT (1UL << PCG_ACCT)
316 /* for encoding cft->private value on file */
319 #define _OOM_TYPE (2)
320 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
321 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
322 #define MEMFILE_ATTR(val) ((val) & 0xffff)
323 /* Used for OOM nofiier */
324 #define OOM_CONTROL (0)
327 * Reclaim flags for mem_cgroup_hierarchical_reclaim
329 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
330 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
331 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
332 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
333 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
334 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
336 static void mem_cgroup_get(struct mem_cgroup *mem);
337 static void mem_cgroup_put(struct mem_cgroup *mem);
338 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
339 static void drain_all_stock_async(void);
341 static struct mem_cgroup_per_zone *
342 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
344 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
347 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
352 static struct mem_cgroup_per_zone *
353 page_cgroup_zoneinfo(struct page_cgroup *pc)
355 struct mem_cgroup *mem = pc->mem_cgroup;
356 int nid = page_cgroup_nid(pc);
357 int zid = page_cgroup_zid(pc);
362 return mem_cgroup_zoneinfo(mem, nid, zid);
365 static struct mem_cgroup_tree_per_zone *
366 soft_limit_tree_node_zone(int nid, int zid)
368 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
371 static struct mem_cgroup_tree_per_zone *
372 soft_limit_tree_from_page(struct page *page)
374 int nid = page_to_nid(page);
375 int zid = page_zonenum(page);
377 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
381 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
382 struct mem_cgroup_per_zone *mz,
383 struct mem_cgroup_tree_per_zone *mctz,
384 unsigned long long new_usage_in_excess)
386 struct rb_node **p = &mctz->rb_root.rb_node;
387 struct rb_node *parent = NULL;
388 struct mem_cgroup_per_zone *mz_node;
393 mz->usage_in_excess = new_usage_in_excess;
394 if (!mz->usage_in_excess)
398 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
400 if (mz->usage_in_excess < mz_node->usage_in_excess)
403 * We can't avoid mem cgroups that are over their soft
404 * limit by the same amount
406 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
409 rb_link_node(&mz->tree_node, parent, p);
410 rb_insert_color(&mz->tree_node, &mctz->rb_root);
415 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
416 struct mem_cgroup_per_zone *mz,
417 struct mem_cgroup_tree_per_zone *mctz)
421 rb_erase(&mz->tree_node, &mctz->rb_root);
426 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
427 struct mem_cgroup_per_zone *mz,
428 struct mem_cgroup_tree_per_zone *mctz)
430 spin_lock(&mctz->lock);
431 __mem_cgroup_remove_exceeded(mem, mz, mctz);
432 spin_unlock(&mctz->lock);
436 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
438 unsigned long long excess;
439 struct mem_cgroup_per_zone *mz;
440 struct mem_cgroup_tree_per_zone *mctz;
441 int nid = page_to_nid(page);
442 int zid = page_zonenum(page);
443 mctz = soft_limit_tree_from_page(page);
446 * Necessary to update all ancestors when hierarchy is used.
447 * because their event counter is not touched.
449 for (; mem; mem = parent_mem_cgroup(mem)) {
450 mz = mem_cgroup_zoneinfo(mem, nid, zid);
451 excess = res_counter_soft_limit_excess(&mem->res);
453 * We have to update the tree if mz is on RB-tree or
454 * mem is over its softlimit.
456 if (excess || mz->on_tree) {
457 spin_lock(&mctz->lock);
458 /* if on-tree, remove it */
460 __mem_cgroup_remove_exceeded(mem, mz, mctz);
462 * Insert again. mz->usage_in_excess will be updated.
463 * If excess is 0, no tree ops.
465 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
466 spin_unlock(&mctz->lock);
471 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
474 struct mem_cgroup_per_zone *mz;
475 struct mem_cgroup_tree_per_zone *mctz;
477 for_each_node_state(node, N_POSSIBLE) {
478 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
479 mz = mem_cgroup_zoneinfo(mem, node, zone);
480 mctz = soft_limit_tree_node_zone(node, zone);
481 mem_cgroup_remove_exceeded(mem, mz, mctz);
486 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
488 return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
491 static struct mem_cgroup_per_zone *
492 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
494 struct rb_node *rightmost = NULL;
495 struct mem_cgroup_per_zone *mz;
499 rightmost = rb_last(&mctz->rb_root);
501 goto done; /* Nothing to reclaim from */
503 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
505 * Remove the node now but someone else can add it back,
506 * we will to add it back at the end of reclaim to its correct
507 * position in the tree.
509 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
510 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
511 !css_tryget(&mz->mem->css))
517 static struct mem_cgroup_per_zone *
518 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
520 struct mem_cgroup_per_zone *mz;
522 spin_lock(&mctz->lock);
523 mz = __mem_cgroup_largest_soft_limit_node(mctz);
524 spin_unlock(&mctz->lock);
528 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
529 enum mem_cgroup_stat_index idx)
534 for_each_possible_cpu(cpu)
535 val += per_cpu(mem->stat->count[idx], cpu);
539 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
543 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
544 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
548 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
551 int val = (charge) ? 1 : -1;
552 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
555 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
556 struct page_cgroup *pc,
559 int val = (charge) ? 1 : -1;
563 if (PageCgroupCache(pc))
564 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
566 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
569 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
571 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
572 __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
577 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
581 struct mem_cgroup_per_zone *mz;
584 for_each_online_node(nid)
585 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
586 mz = mem_cgroup_zoneinfo(mem, nid, zid);
587 total += MEM_CGROUP_ZSTAT(mz, idx);
592 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
596 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
598 return !(val & ((1 << event_mask_shift) - 1));
602 * Check events in order.
605 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
607 /* threshold event is triggered in finer grain than soft limit */
608 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
609 mem_cgroup_threshold(mem);
610 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
611 mem_cgroup_update_tree(mem, page);
615 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
617 return container_of(cgroup_subsys_state(cont,
618 mem_cgroup_subsys_id), struct mem_cgroup,
622 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
625 * mm_update_next_owner() may clear mm->owner to NULL
626 * if it races with swapoff, page migration, etc.
627 * So this can be called with p == NULL.
632 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
633 struct mem_cgroup, css);
636 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
638 struct mem_cgroup *mem = NULL;
643 * Because we have no locks, mm->owner's may be being moved to other
644 * cgroup. We use css_tryget() here even if this looks
645 * pessimistic (rather than adding locks here).
649 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
652 } while (!css_tryget(&mem->css));
658 * Call callback function against all cgroup under hierarchy tree.
660 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
661 int (*func)(struct mem_cgroup *, void *))
663 int found, ret, nextid;
664 struct cgroup_subsys_state *css;
665 struct mem_cgroup *mem;
667 if (!root->use_hierarchy)
668 return (*func)(root, data);
676 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
678 if (css && css_tryget(css))
679 mem = container_of(css, struct mem_cgroup, css);
683 ret = (*func)(mem, data);
687 } while (!ret && css);
692 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
694 return (mem == root_mem_cgroup);
698 * Following LRU functions are allowed to be used without PCG_LOCK.
699 * Operations are called by routine of global LRU independently from memcg.
700 * What we have to take care of here is validness of pc->mem_cgroup.
702 * Changes to pc->mem_cgroup happens when
705 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
706 * It is added to LRU before charge.
707 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
708 * When moving account, the page is not on LRU. It's isolated.
711 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
713 struct page_cgroup *pc;
714 struct mem_cgroup_per_zone *mz;
716 if (mem_cgroup_disabled())
718 pc = lookup_page_cgroup(page);
719 /* can happen while we handle swapcache. */
720 if (!TestClearPageCgroupAcctLRU(pc))
722 VM_BUG_ON(!pc->mem_cgroup);
724 * We don't check PCG_USED bit. It's cleared when the "page" is finally
725 * removed from global LRU.
727 mz = page_cgroup_zoneinfo(pc);
728 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
729 if (mem_cgroup_is_root(pc->mem_cgroup))
731 VM_BUG_ON(list_empty(&pc->lru));
732 list_del_init(&pc->lru);
736 void mem_cgroup_del_lru(struct page *page)
738 mem_cgroup_del_lru_list(page, page_lru(page));
741 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
743 struct mem_cgroup_per_zone *mz;
744 struct page_cgroup *pc;
746 if (mem_cgroup_disabled())
749 pc = lookup_page_cgroup(page);
751 * Used bit is set without atomic ops but after smp_wmb().
752 * For making pc->mem_cgroup visible, insert smp_rmb() here.
755 /* unused or root page is not rotated. */
756 if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
758 mz = page_cgroup_zoneinfo(pc);
759 list_move(&pc->lru, &mz->lists[lru]);
762 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
764 struct page_cgroup *pc;
765 struct mem_cgroup_per_zone *mz;
767 if (mem_cgroup_disabled())
769 pc = lookup_page_cgroup(page);
770 VM_BUG_ON(PageCgroupAcctLRU(pc));
772 * Used bit is set without atomic ops but after smp_wmb().
773 * For making pc->mem_cgroup visible, insert smp_rmb() here.
776 if (!PageCgroupUsed(pc))
779 mz = page_cgroup_zoneinfo(pc);
780 MEM_CGROUP_ZSTAT(mz, lru) += 1;
781 SetPageCgroupAcctLRU(pc);
782 if (mem_cgroup_is_root(pc->mem_cgroup))
784 list_add(&pc->lru, &mz->lists[lru]);
788 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
789 * lru because the page may.be reused after it's fully uncharged (because of
790 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
791 * it again. This function is only used to charge SwapCache. It's done under
792 * lock_page and expected that zone->lru_lock is never held.
794 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
797 struct zone *zone = page_zone(page);
798 struct page_cgroup *pc = lookup_page_cgroup(page);
800 spin_lock_irqsave(&zone->lru_lock, flags);
802 * Forget old LRU when this page_cgroup is *not* used. This Used bit
803 * is guarded by lock_page() because the page is SwapCache.
805 if (!PageCgroupUsed(pc))
806 mem_cgroup_del_lru_list(page, page_lru(page));
807 spin_unlock_irqrestore(&zone->lru_lock, flags);
810 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
813 struct zone *zone = page_zone(page);
814 struct page_cgroup *pc = lookup_page_cgroup(page);
816 spin_lock_irqsave(&zone->lru_lock, flags);
817 /* link when the page is linked to LRU but page_cgroup isn't */
818 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
819 mem_cgroup_add_lru_list(page, page_lru(page));
820 spin_unlock_irqrestore(&zone->lru_lock, flags);
824 void mem_cgroup_move_lists(struct page *page,
825 enum lru_list from, enum lru_list to)
827 if (mem_cgroup_disabled())
829 mem_cgroup_del_lru_list(page, from);
830 mem_cgroup_add_lru_list(page, to);
833 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
836 struct mem_cgroup *curr = NULL;
840 curr = try_get_mem_cgroup_from_mm(task->mm);
846 * We should check use_hierarchy of "mem" not "curr". Because checking
847 * use_hierarchy of "curr" here make this function true if hierarchy is
848 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
849 * hierarchy(even if use_hierarchy is disabled in "mem").
851 if (mem->use_hierarchy)
852 ret = css_is_ancestor(&curr->css, &mem->css);
860 * prev_priority control...this will be used in memory reclaim path.
862 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
866 spin_lock(&mem->reclaim_param_lock);
867 prev_priority = mem->prev_priority;
868 spin_unlock(&mem->reclaim_param_lock);
870 return prev_priority;
873 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
875 spin_lock(&mem->reclaim_param_lock);
876 if (priority < mem->prev_priority)
877 mem->prev_priority = priority;
878 spin_unlock(&mem->reclaim_param_lock);
881 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
883 spin_lock(&mem->reclaim_param_lock);
884 mem->prev_priority = priority;
885 spin_unlock(&mem->reclaim_param_lock);
888 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
890 unsigned long active;
891 unsigned long inactive;
893 unsigned long inactive_ratio;
895 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
896 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
898 gb = (inactive + active) >> (30 - PAGE_SHIFT);
900 inactive_ratio = int_sqrt(10 * gb);
905 present_pages[0] = inactive;
906 present_pages[1] = active;
909 return inactive_ratio;
912 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
914 unsigned long active;
915 unsigned long inactive;
916 unsigned long present_pages[2];
917 unsigned long inactive_ratio;
919 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
921 inactive = present_pages[0];
922 active = present_pages[1];
924 if (inactive * inactive_ratio < active)
930 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
932 unsigned long active;
933 unsigned long inactive;
935 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
936 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
938 return (active > inactive);
941 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
945 int nid = zone->zone_pgdat->node_id;
946 int zid = zone_idx(zone);
947 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
949 return MEM_CGROUP_ZSTAT(mz, lru);
952 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
955 int nid = zone->zone_pgdat->node_id;
956 int zid = zone_idx(zone);
957 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
959 return &mz->reclaim_stat;
962 struct zone_reclaim_stat *
963 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
965 struct page_cgroup *pc;
966 struct mem_cgroup_per_zone *mz;
968 if (mem_cgroup_disabled())
971 pc = lookup_page_cgroup(page);
973 * Used bit is set without atomic ops but after smp_wmb().
974 * For making pc->mem_cgroup visible, insert smp_rmb() here.
977 if (!PageCgroupUsed(pc))
980 mz = page_cgroup_zoneinfo(pc);
984 return &mz->reclaim_stat;
987 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
988 struct list_head *dst,
989 unsigned long *scanned, int order,
990 int mode, struct zone *z,
991 struct mem_cgroup *mem_cont,
992 int active, int file)
994 unsigned long nr_taken = 0;
998 struct list_head *src;
999 struct page_cgroup *pc, *tmp;
1000 int nid = z->zone_pgdat->node_id;
1001 int zid = zone_idx(z);
1002 struct mem_cgroup_per_zone *mz;
1003 int lru = LRU_FILE * file + active;
1007 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1008 src = &mz->lists[lru];
1011 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1012 if (scan >= nr_to_scan)
1016 if (unlikely(!PageCgroupUsed(pc)))
1018 if (unlikely(!PageLRU(page)))
1022 ret = __isolate_lru_page(page, mode, file);
1025 list_move(&page->lru, dst);
1026 mem_cgroup_del_lru(page);
1030 /* we don't affect global LRU but rotate in our LRU */
1031 mem_cgroup_rotate_lru_list(page, page_lru(page));
1042 #define mem_cgroup_from_res_counter(counter, member) \
1043 container_of(counter, struct mem_cgroup, member)
1045 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1047 if (do_swap_account) {
1048 if (res_counter_check_under_limit(&mem->res) &&
1049 res_counter_check_under_limit(&mem->memsw))
1052 if (res_counter_check_under_limit(&mem->res))
1057 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1059 struct cgroup *cgrp = memcg->css.cgroup;
1060 unsigned int swappiness;
1063 if (cgrp->parent == NULL)
1064 return vm_swappiness;
1066 spin_lock(&memcg->reclaim_param_lock);
1067 swappiness = memcg->swappiness;
1068 spin_unlock(&memcg->reclaim_param_lock);
1073 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1081 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1082 * @memcg: The memory cgroup that went over limit
1083 * @p: Task that is going to be killed
1085 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1088 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1090 struct cgroup *task_cgrp;
1091 struct cgroup *mem_cgrp;
1093 * Need a buffer in BSS, can't rely on allocations. The code relies
1094 * on the assumption that OOM is serialized for memory controller.
1095 * If this assumption is broken, revisit this code.
1097 static char memcg_name[PATH_MAX];
1106 mem_cgrp = memcg->css.cgroup;
1107 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1109 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1112 * Unfortunately, we are unable to convert to a useful name
1113 * But we'll still print out the usage information
1120 printk(KERN_INFO "Task in %s killed", memcg_name);
1123 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1131 * Continues from above, so we don't need an KERN_ level
1133 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1136 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1137 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1138 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1139 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1140 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1142 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1143 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1144 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1148 * This function returns the number of memcg under hierarchy tree. Returns
1149 * 1(self count) if no children.
1151 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1154 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1159 * Visit the first child (need not be the first child as per the ordering
1160 * of the cgroup list, since we track last_scanned_child) of @mem and use
1161 * that to reclaim free pages from.
1163 static struct mem_cgroup *
1164 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1166 struct mem_cgroup *ret = NULL;
1167 struct cgroup_subsys_state *css;
1170 if (!root_mem->use_hierarchy) {
1171 css_get(&root_mem->css);
1177 nextid = root_mem->last_scanned_child + 1;
1178 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1180 if (css && css_tryget(css))
1181 ret = container_of(css, struct mem_cgroup, css);
1184 /* Updates scanning parameter */
1185 spin_lock(&root_mem->reclaim_param_lock);
1187 /* this means start scan from ID:1 */
1188 root_mem->last_scanned_child = 0;
1190 root_mem->last_scanned_child = found;
1191 spin_unlock(&root_mem->reclaim_param_lock);
1198 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1199 * we reclaimed from, so that we don't end up penalizing one child extensively
1200 * based on its position in the children list.
1202 * root_mem is the original ancestor that we've been reclaim from.
1204 * We give up and return to the caller when we visit root_mem twice.
1205 * (other groups can be removed while we're walking....)
1207 * If shrink==true, for avoiding to free too much, this returns immedieately.
1209 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1212 unsigned long reclaim_options)
1214 struct mem_cgroup *victim;
1217 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1218 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1219 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1220 unsigned long excess = mem_cgroup_get_excess(root_mem);
1222 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1223 if (root_mem->memsw_is_minimum)
1227 victim = mem_cgroup_select_victim(root_mem);
1228 if (victim == root_mem) {
1231 drain_all_stock_async();
1234 * If we have not been able to reclaim
1235 * anything, it might because there are
1236 * no reclaimable pages under this hierarchy
1238 if (!check_soft || !total) {
1239 css_put(&victim->css);
1243 * We want to do more targetted reclaim.
1244 * excess >> 2 is not to excessive so as to
1245 * reclaim too much, nor too less that we keep
1246 * coming back to reclaim from this cgroup
1248 if (total >= (excess >> 2) ||
1249 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1250 css_put(&victim->css);
1255 if (!mem_cgroup_local_usage(victim)) {
1256 /* this cgroup's local usage == 0 */
1257 css_put(&victim->css);
1260 /* we use swappiness of local cgroup */
1262 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1263 noswap, get_swappiness(victim), zone,
1264 zone->zone_pgdat->node_id);
1266 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1267 noswap, get_swappiness(victim));
1268 css_put(&victim->css);
1270 * At shrinking usage, we can't check we should stop here or
1271 * reclaim more. It's depends on callers. last_scanned_child
1272 * will work enough for keeping fairness under tree.
1278 if (res_counter_check_under_soft_limit(&root_mem->res))
1280 } else if (mem_cgroup_check_under_limit(root_mem))
1286 static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1288 int *val = (int *)data;
1291 * Logically, we can stop scanning immediately when we find
1292 * a memcg is already locked. But condidering unlock ops and
1293 * creation/removal of memcg, scan-all is simple operation.
1295 x = atomic_inc_return(&mem->oom_lock);
1296 *val = max(x, *val);
1300 * Check OOM-Killer is already running under our hierarchy.
1301 * If someone is running, return false.
1303 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1307 mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1309 if (lock_count == 1)
1314 static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1317 * When a new child is created while the hierarchy is under oom,
1318 * mem_cgroup_oom_lock() may not be called. We have to use
1319 * atomic_add_unless() here.
1321 atomic_add_unless(&mem->oom_lock, -1, 0);
1325 static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1327 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1330 static DEFINE_MUTEX(memcg_oom_mutex);
1331 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1333 struct oom_wait_info {
1334 struct mem_cgroup *mem;
1338 static int memcg_oom_wake_function(wait_queue_t *wait,
1339 unsigned mode, int sync, void *arg)
1341 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1342 struct oom_wait_info *oom_wait_info;
1344 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1346 if (oom_wait_info->mem == wake_mem)
1348 /* if no hierarchy, no match */
1349 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1352 * Both of oom_wait_info->mem and wake_mem are stable under us.
1353 * Then we can use css_is_ancestor without taking care of RCU.
1355 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1356 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1360 return autoremove_wake_function(wait, mode, sync, arg);
1363 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1365 /* for filtering, pass "mem" as argument. */
1366 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1369 static void memcg_oom_recover(struct mem_cgroup *mem)
1371 if (mem->oom_kill_disable && atomic_read(&mem->oom_lock))
1372 memcg_wakeup_oom(mem);
1376 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1378 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1380 struct oom_wait_info owait;
1381 bool locked, need_to_kill;
1384 owait.wait.flags = 0;
1385 owait.wait.func = memcg_oom_wake_function;
1386 owait.wait.private = current;
1387 INIT_LIST_HEAD(&owait.wait.task_list);
1388 need_to_kill = true;
1389 /* At first, try to OOM lock hierarchy under mem.*/
1390 mutex_lock(&memcg_oom_mutex);
1391 locked = mem_cgroup_oom_lock(mem);
1393 * Even if signal_pending(), we can't quit charge() loop without
1394 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1395 * under OOM is always welcomed, use TASK_KILLABLE here.
1397 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1398 if (!locked || mem->oom_kill_disable)
1399 need_to_kill = false;
1401 mem_cgroup_oom_notify(mem);
1402 mutex_unlock(&memcg_oom_mutex);
1405 finish_wait(&memcg_oom_waitq, &owait.wait);
1406 mem_cgroup_out_of_memory(mem, mask);
1409 finish_wait(&memcg_oom_waitq, &owait.wait);
1411 mutex_lock(&memcg_oom_mutex);
1412 mem_cgroup_oom_unlock(mem);
1413 memcg_wakeup_oom(mem);
1414 mutex_unlock(&memcg_oom_mutex);
1416 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1418 /* Give chance to dying process */
1419 schedule_timeout(1);
1424 * Currently used to update mapped file statistics, but the routine can be
1425 * generalized to update other statistics as well.
1427 void mem_cgroup_update_file_mapped(struct page *page, int val)
1429 struct mem_cgroup *mem;
1430 struct page_cgroup *pc;
1432 pc = lookup_page_cgroup(page);
1436 lock_page_cgroup(pc);
1437 mem = pc->mem_cgroup;
1438 if (!mem || !PageCgroupUsed(pc))
1442 * Preemption is already disabled. We can use __this_cpu_xxx
1445 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1446 SetPageCgroupFileMapped(pc);
1448 __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1449 ClearPageCgroupFileMapped(pc);
1453 unlock_page_cgroup(pc);
1457 * size of first charge trial. "32" comes from vmscan.c's magic value.
1458 * TODO: maybe necessary to use big numbers in big irons.
1460 #define CHARGE_SIZE (32 * PAGE_SIZE)
1461 struct memcg_stock_pcp {
1462 struct mem_cgroup *cached; /* this never be root cgroup */
1464 struct work_struct work;
1466 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1467 static atomic_t memcg_drain_count;
1470 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1471 * from local stock and true is returned. If the stock is 0 or charges from a
1472 * cgroup which is not current target, returns false. This stock will be
1475 static bool consume_stock(struct mem_cgroup *mem)
1477 struct memcg_stock_pcp *stock;
1480 stock = &get_cpu_var(memcg_stock);
1481 if (mem == stock->cached && stock->charge)
1482 stock->charge -= PAGE_SIZE;
1483 else /* need to call res_counter_charge */
1485 put_cpu_var(memcg_stock);
1490 * Returns stocks cached in percpu to res_counter and reset cached information.
1492 static void drain_stock(struct memcg_stock_pcp *stock)
1494 struct mem_cgroup *old = stock->cached;
1496 if (stock->charge) {
1497 res_counter_uncharge(&old->res, stock->charge);
1498 if (do_swap_account)
1499 res_counter_uncharge(&old->memsw, stock->charge);
1501 stock->cached = NULL;
1506 * This must be called under preempt disabled or must be called by
1507 * a thread which is pinned to local cpu.
1509 static void drain_local_stock(struct work_struct *dummy)
1511 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1516 * Cache charges(val) which is from res_counter, to local per_cpu area.
1517 * This will be consumed by consume_stock() function, later.
1519 static void refill_stock(struct mem_cgroup *mem, int val)
1521 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1523 if (stock->cached != mem) { /* reset if necessary */
1525 stock->cached = mem;
1527 stock->charge += val;
1528 put_cpu_var(memcg_stock);
1532 * Tries to drain stocked charges in other cpus. This function is asynchronous
1533 * and just put a work per cpu for draining localy on each cpu. Caller can
1534 * expects some charges will be back to res_counter later but cannot wait for
1537 static void drain_all_stock_async(void)
1540 /* This function is for scheduling "drain" in asynchronous way.
1541 * The result of "drain" is not directly handled by callers. Then,
1542 * if someone is calling drain, we don't have to call drain more.
1543 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1544 * there is a race. We just do loose check here.
1546 if (atomic_read(&memcg_drain_count))
1548 /* Notify other cpus that system-wide "drain" is running */
1549 atomic_inc(&memcg_drain_count);
1551 for_each_online_cpu(cpu) {
1552 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1553 schedule_work_on(cpu, &stock->work);
1556 atomic_dec(&memcg_drain_count);
1557 /* We don't wait for flush_work */
1560 /* This is a synchronous drain interface. */
1561 static void drain_all_stock_sync(void)
1563 /* called when force_empty is called */
1564 atomic_inc(&memcg_drain_count);
1565 schedule_on_each_cpu(drain_local_stock);
1566 atomic_dec(&memcg_drain_count);
1569 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1570 unsigned long action,
1573 int cpu = (unsigned long)hcpu;
1574 struct memcg_stock_pcp *stock;
1576 if (action != CPU_DEAD)
1578 stock = &per_cpu(memcg_stock, cpu);
1584 * Unlike exported interface, "oom" parameter is added. if oom==true,
1585 * oom-killer can be invoked.
1587 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1588 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1590 struct mem_cgroup *mem, *mem_over_limit;
1591 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1592 struct res_counter *fail_res;
1593 int csize = CHARGE_SIZE;
1596 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1597 * in system level. So, allow to go ahead dying process in addition to
1600 if (unlikely(test_thread_flag(TIF_MEMDIE)
1601 || fatal_signal_pending(current)))
1605 * We always charge the cgroup the mm_struct belongs to.
1606 * The mm_struct's mem_cgroup changes on task migration if the
1607 * thread group leader migrates. It's possible that mm is not
1608 * set, if so charge the init_mm (happens for pagecache usage).
1612 mem = try_get_mem_cgroup_from_mm(mm);
1620 VM_BUG_ON(css_is_removed(&mem->css));
1621 if (mem_cgroup_is_root(mem))
1626 unsigned long flags = 0;
1628 if (consume_stock(mem))
1631 ret = res_counter_charge(&mem->res, csize, &fail_res);
1633 if (!do_swap_account)
1635 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1638 /* mem+swap counter fails */
1639 res_counter_uncharge(&mem->res, csize);
1640 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1641 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1644 /* mem counter fails */
1645 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1648 /* reduce request size and retry */
1649 if (csize > PAGE_SIZE) {
1653 if (!(gfp_mask & __GFP_WAIT))
1656 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1662 * try_to_free_mem_cgroup_pages() might not give us a full
1663 * picture of reclaim. Some pages are reclaimed and might be
1664 * moved to swap cache or just unmapped from the cgroup.
1665 * Check the limit again to see if the reclaim reduced the
1666 * current usage of the cgroup before giving up
1669 if (mem_cgroup_check_under_limit(mem_over_limit))
1672 /* try to avoid oom while someone is moving charge */
1673 if (mc.moving_task && current != mc.moving_task) {
1674 struct mem_cgroup *from, *to;
1675 bool do_continue = false;
1677 * There is a small race that "from" or "to" can be
1678 * freed by rmdir, so we use css_tryget().
1682 if (from && css_tryget(&from->css)) {
1683 if (mem_over_limit->use_hierarchy)
1684 do_continue = css_is_ancestor(
1686 &mem_over_limit->css);
1688 do_continue = (from == mem_over_limit);
1689 css_put(&from->css);
1691 if (!do_continue && to && css_tryget(&to->css)) {
1692 if (mem_over_limit->use_hierarchy)
1693 do_continue = css_is_ancestor(
1695 &mem_over_limit->css);
1697 do_continue = (to == mem_over_limit);
1702 prepare_to_wait(&mc.waitq, &wait,
1703 TASK_INTERRUPTIBLE);
1704 /* moving charge context might have finished. */
1707 finish_wait(&mc.waitq, &wait);
1712 if (!nr_retries--) {
1715 if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) {
1716 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1719 /* When we reach here, current task is dying .*/
1724 if (csize > PAGE_SIZE)
1725 refill_stock(mem, csize - PAGE_SIZE);
1737 * Somemtimes we have to undo a charge we got by try_charge().
1738 * This function is for that and do uncharge, put css's refcnt.
1739 * gotten by try_charge().
1741 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1742 unsigned long count)
1744 if (!mem_cgroup_is_root(mem)) {
1745 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1746 if (do_swap_account)
1747 res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1748 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1749 WARN_ON_ONCE(count > INT_MAX);
1750 __css_put(&mem->css, (int)count);
1752 /* we don't need css_put for root */
1755 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1757 __mem_cgroup_cancel_charge(mem, 1);
1761 * A helper function to get mem_cgroup from ID. must be called under
1762 * rcu_read_lock(). The caller must check css_is_removed() or some if
1763 * it's concern. (dropping refcnt from swap can be called against removed
1766 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1768 struct cgroup_subsys_state *css;
1770 /* ID 0 is unused ID */
1773 css = css_lookup(&mem_cgroup_subsys, id);
1776 return container_of(css, struct mem_cgroup, css);
1779 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1781 struct mem_cgroup *mem = NULL;
1782 struct page_cgroup *pc;
1786 VM_BUG_ON(!PageLocked(page));
1788 pc = lookup_page_cgroup(page);
1789 lock_page_cgroup(pc);
1790 if (PageCgroupUsed(pc)) {
1791 mem = pc->mem_cgroup;
1792 if (mem && !css_tryget(&mem->css))
1794 } else if (PageSwapCache(page)) {
1795 ent.val = page_private(page);
1796 id = lookup_swap_cgroup(ent);
1798 mem = mem_cgroup_lookup(id);
1799 if (mem && !css_tryget(&mem->css))
1803 unlock_page_cgroup(pc);
1808 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1809 * USED state. If already USED, uncharge and return.
1812 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1813 struct page_cgroup *pc,
1814 enum charge_type ctype)
1816 /* try_charge() can return NULL to *memcg, taking care of it. */
1820 lock_page_cgroup(pc);
1821 if (unlikely(PageCgroupUsed(pc))) {
1822 unlock_page_cgroup(pc);
1823 mem_cgroup_cancel_charge(mem);
1827 pc->mem_cgroup = mem;
1829 * We access a page_cgroup asynchronously without lock_page_cgroup().
1830 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1831 * is accessed after testing USED bit. To make pc->mem_cgroup visible
1832 * before USED bit, we need memory barrier here.
1833 * See mem_cgroup_add_lru_list(), etc.
1837 case MEM_CGROUP_CHARGE_TYPE_CACHE:
1838 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1839 SetPageCgroupCache(pc);
1840 SetPageCgroupUsed(pc);
1842 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1843 ClearPageCgroupCache(pc);
1844 SetPageCgroupUsed(pc);
1850 mem_cgroup_charge_statistics(mem, pc, true);
1852 unlock_page_cgroup(pc);
1854 * "charge_statistics" updated event counter. Then, check it.
1855 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1856 * if they exceeds softlimit.
1858 memcg_check_events(mem, pc->page);
1862 * __mem_cgroup_move_account - move account of the page
1863 * @pc: page_cgroup of the page.
1864 * @from: mem_cgroup which the page is moved from.
1865 * @to: mem_cgroup which the page is moved to. @from != @to.
1866 * @uncharge: whether we should call uncharge and css_put against @from.
1868 * The caller must confirm following.
1869 * - page is not on LRU (isolate_page() is useful.)
1870 * - the pc is locked, used, and ->mem_cgroup points to @from.
1872 * This function doesn't do "charge" nor css_get to new cgroup. It should be
1873 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1874 * true, this function does "uncharge" from old cgroup, but it doesn't if
1875 * @uncharge is false, so a caller should do "uncharge".
1878 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1879 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1881 VM_BUG_ON(from == to);
1882 VM_BUG_ON(PageLRU(pc->page));
1883 VM_BUG_ON(!PageCgroupLocked(pc));
1884 VM_BUG_ON(!PageCgroupUsed(pc));
1885 VM_BUG_ON(pc->mem_cgroup != from);
1887 if (PageCgroupFileMapped(pc)) {
1888 /* Update mapped_file data for mem_cgroup */
1890 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1891 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1894 mem_cgroup_charge_statistics(from, pc, false);
1896 /* This is not "cancel", but cancel_charge does all we need. */
1897 mem_cgroup_cancel_charge(from);
1899 /* caller should have done css_get */
1900 pc->mem_cgroup = to;
1901 mem_cgroup_charge_statistics(to, pc, true);
1903 * We charges against "to" which may not have any tasks. Then, "to"
1904 * can be under rmdir(). But in current implementation, caller of
1905 * this function is just force_empty() and move charge, so it's
1906 * garanteed that "to" is never removed. So, we don't check rmdir
1912 * check whether the @pc is valid for moving account and call
1913 * __mem_cgroup_move_account()
1915 static int mem_cgroup_move_account(struct page_cgroup *pc,
1916 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1919 lock_page_cgroup(pc);
1920 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1921 __mem_cgroup_move_account(pc, from, to, uncharge);
1924 unlock_page_cgroup(pc);
1928 memcg_check_events(to, pc->page);
1929 memcg_check_events(from, pc->page);
1934 * move charges to its parent.
1937 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1938 struct mem_cgroup *child,
1941 struct page *page = pc->page;
1942 struct cgroup *cg = child->css.cgroup;
1943 struct cgroup *pcg = cg->parent;
1944 struct mem_cgroup *parent;
1952 if (!get_page_unless_zero(page))
1954 if (isolate_lru_page(page))
1957 parent = mem_cgroup_from_cont(pcg);
1958 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1962 ret = mem_cgroup_move_account(pc, child, parent, true);
1964 mem_cgroup_cancel_charge(parent);
1966 putback_lru_page(page);
1974 * Charge the memory controller for page usage.
1976 * 0 if the charge was successful
1977 * < 0 if the cgroup is over its limit
1979 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1980 gfp_t gfp_mask, enum charge_type ctype,
1981 struct mem_cgroup *memcg)
1983 struct mem_cgroup *mem;
1984 struct page_cgroup *pc;
1987 pc = lookup_page_cgroup(page);
1988 /* can happen at boot */
1994 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1998 __mem_cgroup_commit_charge(mem, pc, ctype);
2002 int mem_cgroup_newpage_charge(struct page *page,
2003 struct mm_struct *mm, gfp_t gfp_mask)
2005 if (mem_cgroup_disabled())
2007 if (PageCompound(page))
2010 * If already mapped, we don't have to account.
2011 * If page cache, page->mapping has address_space.
2012 * But page->mapping may have out-of-use anon_vma pointer,
2013 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2016 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2020 return mem_cgroup_charge_common(page, mm, gfp_mask,
2021 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
2025 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2026 enum charge_type ctype);
2028 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2031 struct mem_cgroup *mem = NULL;
2034 if (mem_cgroup_disabled())
2036 if (PageCompound(page))
2039 * Corner case handling. This is called from add_to_page_cache()
2040 * in usual. But some FS (shmem) precharges this page before calling it
2041 * and call add_to_page_cache() with GFP_NOWAIT.
2043 * For GFP_NOWAIT case, the page may be pre-charged before calling
2044 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2045 * charge twice. (It works but has to pay a bit larger cost.)
2046 * And when the page is SwapCache, it should take swap information
2047 * into account. This is under lock_page() now.
2049 if (!(gfp_mask & __GFP_WAIT)) {
2050 struct page_cgroup *pc;
2053 pc = lookup_page_cgroup(page);
2056 lock_page_cgroup(pc);
2057 if (PageCgroupUsed(pc)) {
2058 unlock_page_cgroup(pc);
2061 unlock_page_cgroup(pc);
2064 if (unlikely(!mm && !mem))
2067 if (page_is_file_cache(page))
2068 return mem_cgroup_charge_common(page, mm, gfp_mask,
2069 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
2072 if (PageSwapCache(page)) {
2073 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2075 __mem_cgroup_commit_charge_swapin(page, mem,
2076 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2078 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2079 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
2085 * While swap-in, try_charge -> commit or cancel, the page is locked.
2086 * And when try_charge() successfully returns, one refcnt to memcg without
2087 * struct page_cgroup is acquired. This refcnt will be consumed by
2088 * "commit()" or removed by "cancel()"
2090 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2092 gfp_t mask, struct mem_cgroup **ptr)
2094 struct mem_cgroup *mem;
2097 if (mem_cgroup_disabled())
2100 if (!do_swap_account)
2103 * A racing thread's fault, or swapoff, may have already updated
2104 * the pte, and even removed page from swap cache: in those cases
2105 * do_swap_page()'s pte_same() test will fail; but there's also a
2106 * KSM case which does need to charge the page.
2108 if (!PageSwapCache(page))
2110 mem = try_get_mem_cgroup_from_page(page);
2114 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2115 /* drop extra refcnt from tryget */
2121 return __mem_cgroup_try_charge(mm, mask, ptr, true);
2125 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2126 enum charge_type ctype)
2128 struct page_cgroup *pc;
2130 if (mem_cgroup_disabled())
2134 cgroup_exclude_rmdir(&ptr->css);
2135 pc = lookup_page_cgroup(page);
2136 mem_cgroup_lru_del_before_commit_swapcache(page);
2137 __mem_cgroup_commit_charge(ptr, pc, ctype);
2138 mem_cgroup_lru_add_after_commit_swapcache(page);
2140 * Now swap is on-memory. This means this page may be
2141 * counted both as mem and swap....double count.
2142 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2143 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2144 * may call delete_from_swap_cache() before reach here.
2146 if (do_swap_account && PageSwapCache(page)) {
2147 swp_entry_t ent = {.val = page_private(page)};
2149 struct mem_cgroup *memcg;
2151 id = swap_cgroup_record(ent, 0);
2153 memcg = mem_cgroup_lookup(id);
2156 * This recorded memcg can be obsolete one. So, avoid
2157 * calling css_tryget
2159 if (!mem_cgroup_is_root(memcg))
2160 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2161 mem_cgroup_swap_statistics(memcg, false);
2162 mem_cgroup_put(memcg);
2167 * At swapin, we may charge account against cgroup which has no tasks.
2168 * So, rmdir()->pre_destroy() can be called while we do this charge.
2169 * In that case, we need to call pre_destroy() again. check it here.
2171 cgroup_release_and_wakeup_rmdir(&ptr->css);
2174 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2176 __mem_cgroup_commit_charge_swapin(page, ptr,
2177 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2180 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2182 if (mem_cgroup_disabled())
2186 mem_cgroup_cancel_charge(mem);
2190 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2192 struct memcg_batch_info *batch = NULL;
2193 bool uncharge_memsw = true;
2194 /* If swapout, usage of swap doesn't decrease */
2195 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2196 uncharge_memsw = false;
2198 batch = ¤t->memcg_batch;
2200 * In usual, we do css_get() when we remember memcg pointer.
2201 * But in this case, we keep res->usage until end of a series of
2202 * uncharges. Then, it's ok to ignore memcg's refcnt.
2207 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2208 * In those cases, all pages freed continously can be expected to be in
2209 * the same cgroup and we have chance to coalesce uncharges.
2210 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2211 * because we want to do uncharge as soon as possible.
2214 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2215 goto direct_uncharge;
2218 * In typical case, batch->memcg == mem. This means we can
2219 * merge a series of uncharges to an uncharge of res_counter.
2220 * If not, we uncharge res_counter ony by one.
2222 if (batch->memcg != mem)
2223 goto direct_uncharge;
2224 /* remember freed charge and uncharge it later */
2225 batch->bytes += PAGE_SIZE;
2227 batch->memsw_bytes += PAGE_SIZE;
2230 res_counter_uncharge(&mem->res, PAGE_SIZE);
2232 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2233 if (unlikely(batch->memcg != mem))
2234 memcg_oom_recover(mem);
2239 * uncharge if !page_mapped(page)
2241 static struct mem_cgroup *
2242 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2244 struct page_cgroup *pc;
2245 struct mem_cgroup *mem = NULL;
2246 struct mem_cgroup_per_zone *mz;
2248 if (mem_cgroup_disabled())
2251 if (PageSwapCache(page))
2255 * Check if our page_cgroup is valid
2257 pc = lookup_page_cgroup(page);
2258 if (unlikely(!pc || !PageCgroupUsed(pc)))
2261 lock_page_cgroup(pc);
2263 mem = pc->mem_cgroup;
2265 if (!PageCgroupUsed(pc))
2269 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2270 case MEM_CGROUP_CHARGE_TYPE_DROP:
2271 /* See mem_cgroup_prepare_migration() */
2272 if (page_mapped(page) || PageCgroupMigration(pc))
2275 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2276 if (!PageAnon(page)) { /* Shared memory */
2277 if (page->mapping && !page_is_file_cache(page))
2279 } else if (page_mapped(page)) /* Anon */
2286 if (!mem_cgroup_is_root(mem))
2287 __do_uncharge(mem, ctype);
2288 if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2289 mem_cgroup_swap_statistics(mem, true);
2290 mem_cgroup_charge_statistics(mem, pc, false);
2292 ClearPageCgroupUsed(pc);
2294 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2295 * freed from LRU. This is safe because uncharged page is expected not
2296 * to be reused (freed soon). Exception is SwapCache, it's handled by
2297 * special functions.
2300 mz = page_cgroup_zoneinfo(pc);
2301 unlock_page_cgroup(pc);
2303 memcg_check_events(mem, page);
2304 /* at swapout, this memcg will be accessed to record to swap */
2305 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2311 unlock_page_cgroup(pc);
2315 void mem_cgroup_uncharge_page(struct page *page)
2318 if (page_mapped(page))
2320 if (page->mapping && !PageAnon(page))
2322 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2325 void mem_cgroup_uncharge_cache_page(struct page *page)
2327 VM_BUG_ON(page_mapped(page));
2328 VM_BUG_ON(page->mapping);
2329 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2333 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2334 * In that cases, pages are freed continuously and we can expect pages
2335 * are in the same memcg. All these calls itself limits the number of
2336 * pages freed at once, then uncharge_start/end() is called properly.
2337 * This may be called prural(2) times in a context,
2340 void mem_cgroup_uncharge_start(void)
2342 current->memcg_batch.do_batch++;
2343 /* We can do nest. */
2344 if (current->memcg_batch.do_batch == 1) {
2345 current->memcg_batch.memcg = NULL;
2346 current->memcg_batch.bytes = 0;
2347 current->memcg_batch.memsw_bytes = 0;
2351 void mem_cgroup_uncharge_end(void)
2353 struct memcg_batch_info *batch = ¤t->memcg_batch;
2355 if (!batch->do_batch)
2359 if (batch->do_batch) /* If stacked, do nothing. */
2365 * This "batch->memcg" is valid without any css_get/put etc...
2366 * bacause we hide charges behind us.
2369 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2370 if (batch->memsw_bytes)
2371 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2372 memcg_oom_recover(batch->memcg);
2373 /* forget this pointer (for sanity check) */
2374 batch->memcg = NULL;
2379 * called after __delete_from_swap_cache() and drop "page" account.
2380 * memcg information is recorded to swap_cgroup of "ent"
2383 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2385 struct mem_cgroup *memcg;
2386 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2388 if (!swapout) /* this was a swap cache but the swap is unused ! */
2389 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2391 memcg = __mem_cgroup_uncharge_common(page, ctype);
2393 /* record memcg information */
2394 if (do_swap_account && swapout && memcg) {
2395 swap_cgroup_record(ent, css_id(&memcg->css));
2396 mem_cgroup_get(memcg);
2398 if (swapout && memcg)
2399 css_put(&memcg->css);
2403 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2405 * called from swap_entry_free(). remove record in swap_cgroup and
2406 * uncharge "memsw" account.
2408 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2410 struct mem_cgroup *memcg;
2413 if (!do_swap_account)
2416 id = swap_cgroup_record(ent, 0);
2418 memcg = mem_cgroup_lookup(id);
2421 * We uncharge this because swap is freed.
2422 * This memcg can be obsolete one. We avoid calling css_tryget
2424 if (!mem_cgroup_is_root(memcg))
2425 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2426 mem_cgroup_swap_statistics(memcg, false);
2427 mem_cgroup_put(memcg);
2433 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2434 * @entry: swap entry to be moved
2435 * @from: mem_cgroup which the entry is moved from
2436 * @to: mem_cgroup which the entry is moved to
2437 * @need_fixup: whether we should fixup res_counters and refcounts.
2439 * It succeeds only when the swap_cgroup's record for this entry is the same
2440 * as the mem_cgroup's id of @from.
2442 * Returns 0 on success, -EINVAL on failure.
2444 * The caller must have charged to @to, IOW, called res_counter_charge() about
2445 * both res and memsw, and called css_get().
2447 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2448 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2450 unsigned short old_id, new_id;
2452 old_id = css_id(&from->css);
2453 new_id = css_id(&to->css);
2455 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2456 mem_cgroup_swap_statistics(from, false);
2457 mem_cgroup_swap_statistics(to, true);
2459 * This function is only called from task migration context now.
2460 * It postpones res_counter and refcount handling till the end
2461 * of task migration(mem_cgroup_clear_mc()) for performance
2462 * improvement. But we cannot postpone mem_cgroup_get(to)
2463 * because if the process that has been moved to @to does
2464 * swap-in, the refcount of @to might be decreased to 0.
2468 if (!mem_cgroup_is_root(from))
2469 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2470 mem_cgroup_put(from);
2472 * we charged both to->res and to->memsw, so we should
2475 if (!mem_cgroup_is_root(to))
2476 res_counter_uncharge(&to->res, PAGE_SIZE);
2484 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2485 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2492 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2495 int mem_cgroup_prepare_migration(struct page *page,
2496 struct page *newpage, struct mem_cgroup **ptr)
2498 struct page_cgroup *pc;
2499 struct mem_cgroup *mem = NULL;
2500 enum charge_type ctype;
2503 if (mem_cgroup_disabled())
2506 pc = lookup_page_cgroup(page);
2507 lock_page_cgroup(pc);
2508 if (PageCgroupUsed(pc)) {
2509 mem = pc->mem_cgroup;
2512 * At migrating an anonymous page, its mapcount goes down
2513 * to 0 and uncharge() will be called. But, even if it's fully
2514 * unmapped, migration may fail and this page has to be
2515 * charged again. We set MIGRATION flag here and delay uncharge
2516 * until end_migration() is called
2518 * Corner Case Thinking
2520 * When the old page was mapped as Anon and it's unmap-and-freed
2521 * while migration was ongoing.
2522 * If unmap finds the old page, uncharge() of it will be delayed
2523 * until end_migration(). If unmap finds a new page, it's
2524 * uncharged when it make mapcount to be 1->0. If unmap code
2525 * finds swap_migration_entry, the new page will not be mapped
2526 * and end_migration() will find it(mapcount==0).
2529 * When the old page was mapped but migraion fails, the kernel
2530 * remaps it. A charge for it is kept by MIGRATION flag even
2531 * if mapcount goes down to 0. We can do remap successfully
2532 * without charging it again.
2535 * The "old" page is under lock_page() until the end of
2536 * migration, so, the old page itself will not be swapped-out.
2537 * If the new page is swapped out before end_migraton, our
2538 * hook to usual swap-out path will catch the event.
2541 SetPageCgroupMigration(pc);
2543 unlock_page_cgroup(pc);
2545 * If the page is not charged at this point,
2552 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2553 css_put(&mem->css);/* drop extra refcnt */
2554 if (ret || *ptr == NULL) {
2555 if (PageAnon(page)) {
2556 lock_page_cgroup(pc);
2557 ClearPageCgroupMigration(pc);
2558 unlock_page_cgroup(pc);
2560 * The old page may be fully unmapped while we kept it.
2562 mem_cgroup_uncharge_page(page);
2567 * We charge new page before it's used/mapped. So, even if unlock_page()
2568 * is called before end_migration, we can catch all events on this new
2569 * page. In the case new page is migrated but not remapped, new page's
2570 * mapcount will be finally 0 and we call uncharge in end_migration().
2572 pc = lookup_page_cgroup(newpage);
2574 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2575 else if (page_is_file_cache(page))
2576 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2578 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2579 __mem_cgroup_commit_charge(mem, pc, ctype);
2583 /* remove redundant charge if migration failed*/
2584 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2585 struct page *oldpage, struct page *newpage)
2587 struct page *used, *unused;
2588 struct page_cgroup *pc;
2592 /* blocks rmdir() */
2593 cgroup_exclude_rmdir(&mem->css);
2594 /* at migration success, oldpage->mapping is NULL. */
2595 if (oldpage->mapping) {
2603 * We disallowed uncharge of pages under migration because mapcount
2604 * of the page goes down to zero, temporarly.
2605 * Clear the flag and check the page should be charged.
2607 pc = lookup_page_cgroup(oldpage);
2608 lock_page_cgroup(pc);
2609 ClearPageCgroupMigration(pc);
2610 unlock_page_cgroup(pc);
2612 if (unused != oldpage)
2613 pc = lookup_page_cgroup(unused);
2614 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2616 pc = lookup_page_cgroup(used);
2618 * If a page is a file cache, radix-tree replacement is very atomic
2619 * and we can skip this check. When it was an Anon page, its mapcount
2620 * goes down to 0. But because we added MIGRATION flage, it's not
2621 * uncharged yet. There are several case but page->mapcount check
2622 * and USED bit check in mem_cgroup_uncharge_page() will do enough
2623 * check. (see prepare_charge() also)
2626 mem_cgroup_uncharge_page(used);
2628 * At migration, we may charge account against cgroup which has no
2630 * So, rmdir()->pre_destroy() can be called while we do this charge.
2631 * In that case, we need to call pre_destroy() again. check it here.
2633 cgroup_release_and_wakeup_rmdir(&mem->css);
2637 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2638 * Calling hierarchical_reclaim is not enough because we should update
2639 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2640 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2641 * not from the memcg which this page would be charged to.
2642 * try_charge_swapin does all of these works properly.
2644 int mem_cgroup_shmem_charge_fallback(struct page *page,
2645 struct mm_struct *mm,
2648 struct mem_cgroup *mem = NULL;
2651 if (mem_cgroup_disabled())
2654 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2656 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2661 static DEFINE_MUTEX(set_limit_mutex);
2663 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2664 unsigned long long val)
2667 u64 memswlimit, memlimit;
2669 int children = mem_cgroup_count_children(memcg);
2670 u64 curusage, oldusage;
2674 * For keeping hierarchical_reclaim simple, how long we should retry
2675 * is depends on callers. We set our retry-count to be function
2676 * of # of children which we should visit in this loop.
2678 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2680 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2683 while (retry_count) {
2684 if (signal_pending(current)) {
2689 * Rather than hide all in some function, I do this in
2690 * open coded manner. You see what this really does.
2691 * We have to guarantee mem->res.limit < mem->memsw.limit.
2693 mutex_lock(&set_limit_mutex);
2694 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2695 if (memswlimit < val) {
2697 mutex_unlock(&set_limit_mutex);
2701 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2705 ret = res_counter_set_limit(&memcg->res, val);
2707 if (memswlimit == val)
2708 memcg->memsw_is_minimum = true;
2710 memcg->memsw_is_minimum = false;
2712 mutex_unlock(&set_limit_mutex);
2717 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2718 MEM_CGROUP_RECLAIM_SHRINK);
2719 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2720 /* Usage is reduced ? */
2721 if (curusage >= oldusage)
2724 oldusage = curusage;
2726 if (!ret && enlarge)
2727 memcg_oom_recover(memcg);
2732 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2733 unsigned long long val)
2736 u64 memlimit, memswlimit, oldusage, curusage;
2737 int children = mem_cgroup_count_children(memcg);
2741 /* see mem_cgroup_resize_res_limit */
2742 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2743 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2744 while (retry_count) {
2745 if (signal_pending(current)) {
2750 * Rather than hide all in some function, I do this in
2751 * open coded manner. You see what this really does.
2752 * We have to guarantee mem->res.limit < mem->memsw.limit.
2754 mutex_lock(&set_limit_mutex);
2755 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2756 if (memlimit > val) {
2758 mutex_unlock(&set_limit_mutex);
2761 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2762 if (memswlimit < val)
2764 ret = res_counter_set_limit(&memcg->memsw, val);
2766 if (memlimit == val)
2767 memcg->memsw_is_minimum = true;
2769 memcg->memsw_is_minimum = false;
2771 mutex_unlock(&set_limit_mutex);
2776 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2777 MEM_CGROUP_RECLAIM_NOSWAP |
2778 MEM_CGROUP_RECLAIM_SHRINK);
2779 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2780 /* Usage is reduced ? */
2781 if (curusage >= oldusage)
2784 oldusage = curusage;
2786 if (!ret && enlarge)
2787 memcg_oom_recover(memcg);
2791 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2792 gfp_t gfp_mask, int nid,
2795 unsigned long nr_reclaimed = 0;
2796 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2797 unsigned long reclaimed;
2799 struct mem_cgroup_tree_per_zone *mctz;
2800 unsigned long long excess;
2805 mctz = soft_limit_tree_node_zone(nid, zid);
2807 * This loop can run a while, specially if mem_cgroup's continuously
2808 * keep exceeding their soft limit and putting the system under
2815 mz = mem_cgroup_largest_soft_limit_node(mctz);
2819 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2821 MEM_CGROUP_RECLAIM_SOFT);
2822 nr_reclaimed += reclaimed;
2823 spin_lock(&mctz->lock);
2826 * If we failed to reclaim anything from this memory cgroup
2827 * it is time to move on to the next cgroup
2833 * Loop until we find yet another one.
2835 * By the time we get the soft_limit lock
2836 * again, someone might have aded the
2837 * group back on the RB tree. Iterate to
2838 * make sure we get a different mem.
2839 * mem_cgroup_largest_soft_limit_node returns
2840 * NULL if no other cgroup is present on
2844 __mem_cgroup_largest_soft_limit_node(mctz);
2845 if (next_mz == mz) {
2846 css_put(&next_mz->mem->css);
2848 } else /* next_mz == NULL or other memcg */
2852 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2853 excess = res_counter_soft_limit_excess(&mz->mem->res);
2855 * One school of thought says that we should not add
2856 * back the node to the tree if reclaim returns 0.
2857 * But our reclaim could return 0, simply because due
2858 * to priority we are exposing a smaller subset of
2859 * memory to reclaim from. Consider this as a longer
2862 /* If excess == 0, no tree ops */
2863 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2864 spin_unlock(&mctz->lock);
2865 css_put(&mz->mem->css);
2868 * Could not reclaim anything and there are no more
2869 * mem cgroups to try or we seem to be looping without
2870 * reclaiming anything.
2872 if (!nr_reclaimed &&
2874 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2876 } while (!nr_reclaimed);
2878 css_put(&next_mz->mem->css);
2879 return nr_reclaimed;
2883 * This routine traverse page_cgroup in given list and drop them all.
2884 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2886 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2887 int node, int zid, enum lru_list lru)
2890 struct mem_cgroup_per_zone *mz;
2891 struct page_cgroup *pc, *busy;
2892 unsigned long flags, loop;
2893 struct list_head *list;
2896 zone = &NODE_DATA(node)->node_zones[zid];
2897 mz = mem_cgroup_zoneinfo(mem, node, zid);
2898 list = &mz->lists[lru];
2900 loop = MEM_CGROUP_ZSTAT(mz, lru);
2901 /* give some margin against EBUSY etc...*/
2906 spin_lock_irqsave(&zone->lru_lock, flags);
2907 if (list_empty(list)) {
2908 spin_unlock_irqrestore(&zone->lru_lock, flags);
2911 pc = list_entry(list->prev, struct page_cgroup, lru);
2913 list_move(&pc->lru, list);
2915 spin_unlock_irqrestore(&zone->lru_lock, flags);
2918 spin_unlock_irqrestore(&zone->lru_lock, flags);
2920 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2924 if (ret == -EBUSY || ret == -EINVAL) {
2925 /* found lock contention or "pc" is obsolete. */
2932 if (!ret && !list_empty(list))
2938 * make mem_cgroup's charge to be 0 if there is no task.
2939 * This enables deleting this mem_cgroup.
2941 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2944 int node, zid, shrink;
2945 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2946 struct cgroup *cgrp = mem->css.cgroup;
2951 /* should free all ? */
2957 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2960 if (signal_pending(current))
2962 /* This is for making all *used* pages to be on LRU. */
2963 lru_add_drain_all();
2964 drain_all_stock_sync();
2966 for_each_node_state(node, N_HIGH_MEMORY) {
2967 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2970 ret = mem_cgroup_force_empty_list(mem,
2979 memcg_oom_recover(mem);
2980 /* it seems parent cgroup doesn't have enough mem */
2984 /* "ret" should also be checked to ensure all lists are empty. */
2985 } while (mem->res.usage > 0 || ret);
2991 /* returns EBUSY if there is a task or if we come here twice. */
2992 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2996 /* we call try-to-free pages for make this cgroup empty */
2997 lru_add_drain_all();
2998 /* try to free all pages in this cgroup */
3000 while (nr_retries && mem->res.usage > 0) {
3003 if (signal_pending(current)) {
3007 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3008 false, get_swappiness(mem));
3011 /* maybe some writeback is necessary */
3012 congestion_wait(BLK_RW_ASYNC, HZ/10);
3017 /* try move_account...there may be some *locked* pages. */
3021 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3023 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3027 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3029 return mem_cgroup_from_cont(cont)->use_hierarchy;
3032 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3036 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3037 struct cgroup *parent = cont->parent;
3038 struct mem_cgroup *parent_mem = NULL;
3041 parent_mem = mem_cgroup_from_cont(parent);
3045 * If parent's use_hierarchy is set, we can't make any modifications
3046 * in the child subtrees. If it is unset, then the change can
3047 * occur, provided the current cgroup has no children.
3049 * For the root cgroup, parent_mem is NULL, we allow value to be
3050 * set if there are no children.
3052 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3053 (val == 1 || val == 0)) {
3054 if (list_empty(&cont->children))
3055 mem->use_hierarchy = val;
3065 struct mem_cgroup_idx_data {
3067 enum mem_cgroup_stat_index idx;
3071 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
3073 struct mem_cgroup_idx_data *d = data;
3074 d->val += mem_cgroup_read_stat(mem, d->idx);
3079 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3080 enum mem_cgroup_stat_index idx, s64 *val)
3082 struct mem_cgroup_idx_data d;
3085 mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
3089 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3093 if (!mem_cgroup_is_root(mem)) {
3095 return res_counter_read_u64(&mem->res, RES_USAGE);
3097 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3100 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
3102 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
3106 mem_cgroup_get_recursive_idx_stat(mem,
3107 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
3111 return val << PAGE_SHIFT;
3114 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3116 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3120 type = MEMFILE_TYPE(cft->private);
3121 name = MEMFILE_ATTR(cft->private);
3124 if (name == RES_USAGE)
3125 val = mem_cgroup_usage(mem, false);
3127 val = res_counter_read_u64(&mem->res, name);
3130 if (name == RES_USAGE)
3131 val = mem_cgroup_usage(mem, true);
3133 val = res_counter_read_u64(&mem->memsw, name);
3142 * The user of this function is...
3145 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3148 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3150 unsigned long long val;
3153 type = MEMFILE_TYPE(cft->private);
3154 name = MEMFILE_ATTR(cft->private);
3157 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3161 /* This function does all necessary parse...reuse it */
3162 ret = res_counter_memparse_write_strategy(buffer, &val);
3166 ret = mem_cgroup_resize_limit(memcg, val);
3168 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3170 case RES_SOFT_LIMIT:
3171 ret = res_counter_memparse_write_strategy(buffer, &val);
3175 * For memsw, soft limits are hard to implement in terms
3176 * of semantics, for now, we support soft limits for
3177 * control without swap
3180 ret = res_counter_set_soft_limit(&memcg->res, val);
3185 ret = -EINVAL; /* should be BUG() ? */
3191 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3192 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3194 struct cgroup *cgroup;
3195 unsigned long long min_limit, min_memsw_limit, tmp;
3197 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3198 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3199 cgroup = memcg->css.cgroup;
3200 if (!memcg->use_hierarchy)
3203 while (cgroup->parent) {
3204 cgroup = cgroup->parent;
3205 memcg = mem_cgroup_from_cont(cgroup);
3206 if (!memcg->use_hierarchy)
3208 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3209 min_limit = min(min_limit, tmp);
3210 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3211 min_memsw_limit = min(min_memsw_limit, tmp);
3214 *mem_limit = min_limit;
3215 *memsw_limit = min_memsw_limit;
3219 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3221 struct mem_cgroup *mem;
3224 mem = mem_cgroup_from_cont(cont);
3225 type = MEMFILE_TYPE(event);
3226 name = MEMFILE_ATTR(event);
3230 res_counter_reset_max(&mem->res);
3232 res_counter_reset_max(&mem->memsw);
3236 res_counter_reset_failcnt(&mem->res);
3238 res_counter_reset_failcnt(&mem->memsw);
3245 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3248 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3252 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3253 struct cftype *cft, u64 val)
3255 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3257 if (val >= (1 << NR_MOVE_TYPE))
3260 * We check this value several times in both in can_attach() and
3261 * attach(), so we need cgroup lock to prevent this value from being
3265 mem->move_charge_at_immigrate = val;
3271 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3272 struct cftype *cft, u64 val)
3279 /* For read statistics */
3295 struct mcs_total_stat {
3296 s64 stat[NR_MCS_STAT];
3302 } memcg_stat_strings[NR_MCS_STAT] = {
3303 {"cache", "total_cache"},
3304 {"rss", "total_rss"},
3305 {"mapped_file", "total_mapped_file"},
3306 {"pgpgin", "total_pgpgin"},
3307 {"pgpgout", "total_pgpgout"},
3308 {"swap", "total_swap"},
3309 {"inactive_anon", "total_inactive_anon"},
3310 {"active_anon", "total_active_anon"},
3311 {"inactive_file", "total_inactive_file"},
3312 {"active_file", "total_active_file"},
3313 {"unevictable", "total_unevictable"}
3317 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3319 struct mcs_total_stat *s = data;
3323 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3324 s->stat[MCS_CACHE] += val * PAGE_SIZE;
3325 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3326 s->stat[MCS_RSS] += val * PAGE_SIZE;
3327 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3328 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3329 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3330 s->stat[MCS_PGPGIN] += val;
3331 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3332 s->stat[MCS_PGPGOUT] += val;
3333 if (do_swap_account) {
3334 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3335 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3339 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3340 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3341 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3342 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3343 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3344 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3345 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3346 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3347 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3348 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3353 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3355 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3358 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3359 struct cgroup_map_cb *cb)
3361 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3362 struct mcs_total_stat mystat;
3365 memset(&mystat, 0, sizeof(mystat));
3366 mem_cgroup_get_local_stat(mem_cont, &mystat);
3368 for (i = 0; i < NR_MCS_STAT; i++) {
3369 if (i == MCS_SWAP && !do_swap_account)
3371 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3374 /* Hierarchical information */
3376 unsigned long long limit, memsw_limit;
3377 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3378 cb->fill(cb, "hierarchical_memory_limit", limit);
3379 if (do_swap_account)
3380 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3383 memset(&mystat, 0, sizeof(mystat));
3384 mem_cgroup_get_total_stat(mem_cont, &mystat);
3385 for (i = 0; i < NR_MCS_STAT; i++) {
3386 if (i == MCS_SWAP && !do_swap_account)
3388 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3391 #ifdef CONFIG_DEBUG_VM
3392 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3396 struct mem_cgroup_per_zone *mz;
3397 unsigned long recent_rotated[2] = {0, 0};
3398 unsigned long recent_scanned[2] = {0, 0};
3400 for_each_online_node(nid)
3401 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3402 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3404 recent_rotated[0] +=
3405 mz->reclaim_stat.recent_rotated[0];
3406 recent_rotated[1] +=
3407 mz->reclaim_stat.recent_rotated[1];
3408 recent_scanned[0] +=
3409 mz->reclaim_stat.recent_scanned[0];
3410 recent_scanned[1] +=
3411 mz->reclaim_stat.recent_scanned[1];
3413 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3414 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3415 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3416 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3423 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3425 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3427 return get_swappiness(memcg);
3430 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3433 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3434 struct mem_cgroup *parent;
3439 if (cgrp->parent == NULL)
3442 parent = mem_cgroup_from_cont(cgrp->parent);
3446 /* If under hierarchy, only empty-root can set this value */
3447 if ((parent->use_hierarchy) ||
3448 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3453 spin_lock(&memcg->reclaim_param_lock);
3454 memcg->swappiness = val;
3455 spin_unlock(&memcg->reclaim_param_lock);
3462 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3464 struct mem_cgroup_threshold_ary *t;
3470 t = rcu_dereference(memcg->thresholds);
3472 t = rcu_dereference(memcg->memsw_thresholds);
3477 usage = mem_cgroup_usage(memcg, swap);
3480 * current_threshold points to threshold just below usage.
3481 * If it's not true, a threshold was crossed after last
3482 * call of __mem_cgroup_threshold().
3484 i = t->current_threshold;
3487 * Iterate backward over array of thresholds starting from
3488 * current_threshold and check if a threshold is crossed.
3489 * If none of thresholds below usage is crossed, we read
3490 * only one element of the array here.
3492 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3493 eventfd_signal(t->entries[i].eventfd, 1);
3495 /* i = current_threshold + 1 */
3499 * Iterate forward over array of thresholds starting from
3500 * current_threshold+1 and check if a threshold is crossed.
3501 * If none of thresholds above usage is crossed, we read
3502 * only one element of the array here.
3504 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3505 eventfd_signal(t->entries[i].eventfd, 1);
3507 /* Update current_threshold */
3508 t->current_threshold = i - 1;
3513 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3515 __mem_cgroup_threshold(memcg, false);
3516 if (do_swap_account)
3517 __mem_cgroup_threshold(memcg, true);
3520 static int compare_thresholds(const void *a, const void *b)
3522 const struct mem_cgroup_threshold *_a = a;
3523 const struct mem_cgroup_threshold *_b = b;
3525 return _a->threshold - _b->threshold;
3528 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem, void *data)
3530 struct mem_cgroup_eventfd_list *ev;
3532 list_for_each_entry(ev, &mem->oom_notify, list)
3533 eventfd_signal(ev->eventfd, 1);
3537 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3539 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_notify_cb);
3542 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3543 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3545 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3546 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3547 int type = MEMFILE_TYPE(cft->private);
3548 u64 threshold, usage;
3552 ret = res_counter_memparse_write_strategy(args, &threshold);
3556 mutex_lock(&memcg->thresholds_lock);
3558 thresholds = memcg->thresholds;
3559 else if (type == _MEMSWAP)
3560 thresholds = memcg->memsw_thresholds;
3564 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3566 /* Check if a threshold crossed before adding a new one */
3568 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3571 size = thresholds->size + 1;
3575 /* Allocate memory for new array of thresholds */
3576 thresholds_new = kmalloc(sizeof(*thresholds_new) +
3577 size * sizeof(struct mem_cgroup_threshold),
3579 if (!thresholds_new) {
3583 thresholds_new->size = size;
3585 /* Copy thresholds (if any) to new array */
3587 memcpy(thresholds_new->entries, thresholds->entries,
3589 sizeof(struct mem_cgroup_threshold));
3590 /* Add new threshold */
3591 thresholds_new->entries[size - 1].eventfd = eventfd;
3592 thresholds_new->entries[size - 1].threshold = threshold;
3594 /* Sort thresholds. Registering of new threshold isn't time-critical */
3595 sort(thresholds_new->entries, size,
3596 sizeof(struct mem_cgroup_threshold),
3597 compare_thresholds, NULL);
3599 /* Find current threshold */
3600 thresholds_new->current_threshold = -1;
3601 for (i = 0; i < size; i++) {
3602 if (thresholds_new->entries[i].threshold < usage) {
3604 * thresholds_new->current_threshold will not be used
3605 * until rcu_assign_pointer(), so it's safe to increment
3608 ++thresholds_new->current_threshold;
3613 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3615 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3617 /* To be sure that nobody uses thresholds */
3621 * Free old preallocated buffer and use thresholds as new
3622 * preallocated buffer.
3625 kfree(memcg->__thresholds);
3626 memcg->__thresholds = thresholds;
3628 kfree(memcg->__memsw_thresholds);
3629 memcg->__memsw_thresholds = thresholds;
3632 mutex_unlock(&memcg->thresholds_lock);
3637 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3638 struct cftype *cft, struct eventfd_ctx *eventfd)
3640 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3641 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3642 int type = MEMFILE_TYPE(cft->private);
3647 mutex_lock(&memcg->thresholds_lock);
3649 thresholds = memcg->thresholds;
3650 else if (type == _MEMSWAP)
3651 thresholds = memcg->memsw_thresholds;
3656 * Something went wrong if we trying to unregister a threshold
3657 * if we don't have thresholds
3659 BUG_ON(!thresholds);
3661 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3663 /* Check if a threshold crossed before removing */
3664 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3666 /* Calculate new number of threshold */
3667 for (i = 0; i < thresholds->size; i++) {
3668 if (thresholds->entries[i].eventfd != eventfd)
3672 /* Use preallocated buffer for new array of thresholds */
3674 thresholds_new = memcg->__thresholds;
3676 thresholds_new = memcg->__memsw_thresholds;
3678 /* Set thresholds array to NULL if we don't have thresholds */
3680 kfree(thresholds_new);
3681 thresholds_new = NULL;
3685 thresholds_new->size = size;
3687 /* Copy thresholds and find current threshold */
3688 thresholds_new->current_threshold = -1;
3689 for (i = 0, j = 0; i < thresholds->size; i++) {
3690 if (thresholds->entries[i].eventfd == eventfd)
3693 thresholds_new->entries[j] = thresholds->entries[i];
3694 if (thresholds_new->entries[j].threshold < usage) {
3696 * thresholds_new->current_threshold will not be used
3697 * until rcu_assign_pointer(), so it's safe to increment
3700 ++thresholds_new->current_threshold;
3706 /* Swap thresholds array and preallocated buffer */
3708 memcg->__thresholds = thresholds;
3709 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3711 memcg->__memsw_thresholds = thresholds;
3712 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3715 /* To be sure that nobody uses thresholds */
3718 mutex_unlock(&memcg->thresholds_lock);
3721 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3722 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3724 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3725 struct mem_cgroup_eventfd_list *event;
3726 int type = MEMFILE_TYPE(cft->private);
3728 BUG_ON(type != _OOM_TYPE);
3729 event = kmalloc(sizeof(*event), GFP_KERNEL);
3733 mutex_lock(&memcg_oom_mutex);
3735 event->eventfd = eventfd;
3736 list_add(&event->list, &memcg->oom_notify);
3738 /* already in OOM ? */
3739 if (atomic_read(&memcg->oom_lock))
3740 eventfd_signal(eventfd, 1);
3741 mutex_unlock(&memcg_oom_mutex);
3746 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3747 struct cftype *cft, struct eventfd_ctx *eventfd)
3749 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3750 struct mem_cgroup_eventfd_list *ev, *tmp;
3751 int type = MEMFILE_TYPE(cft->private);
3753 BUG_ON(type != _OOM_TYPE);
3755 mutex_lock(&memcg_oom_mutex);
3757 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3758 if (ev->eventfd == eventfd) {
3759 list_del(&ev->list);
3764 mutex_unlock(&memcg_oom_mutex);
3767 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
3768 struct cftype *cft, struct cgroup_map_cb *cb)
3770 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3772 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
3774 if (atomic_read(&mem->oom_lock))
3775 cb->fill(cb, "under_oom", 1);
3777 cb->fill(cb, "under_oom", 0);
3783 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
3784 struct cftype *cft, u64 val)
3786 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3787 struct mem_cgroup *parent;
3789 /* cannot set to root cgroup and only 0 and 1 are allowed */
3790 if (!cgrp->parent || !((val == 0) || (val == 1)))
3793 parent = mem_cgroup_from_cont(cgrp->parent);
3796 /* oom-kill-disable is a flag for subhierarchy. */
3797 if ((parent->use_hierarchy) ||
3798 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
3802 mem->oom_kill_disable = val;
3807 static struct cftype mem_cgroup_files[] = {
3809 .name = "usage_in_bytes",
3810 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3811 .read_u64 = mem_cgroup_read,
3812 .register_event = mem_cgroup_usage_register_event,
3813 .unregister_event = mem_cgroup_usage_unregister_event,
3816 .name = "max_usage_in_bytes",
3817 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3818 .trigger = mem_cgroup_reset,
3819 .read_u64 = mem_cgroup_read,
3822 .name = "limit_in_bytes",
3823 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3824 .write_string = mem_cgroup_write,
3825 .read_u64 = mem_cgroup_read,
3828 .name = "soft_limit_in_bytes",
3829 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3830 .write_string = mem_cgroup_write,
3831 .read_u64 = mem_cgroup_read,
3835 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3836 .trigger = mem_cgroup_reset,
3837 .read_u64 = mem_cgroup_read,
3841 .read_map = mem_control_stat_show,
3844 .name = "force_empty",
3845 .trigger = mem_cgroup_force_empty_write,
3848 .name = "use_hierarchy",
3849 .write_u64 = mem_cgroup_hierarchy_write,
3850 .read_u64 = mem_cgroup_hierarchy_read,
3853 .name = "swappiness",
3854 .read_u64 = mem_cgroup_swappiness_read,
3855 .write_u64 = mem_cgroup_swappiness_write,
3858 .name = "move_charge_at_immigrate",
3859 .read_u64 = mem_cgroup_move_charge_read,
3860 .write_u64 = mem_cgroup_move_charge_write,
3863 .name = "oom_control",
3864 .read_map = mem_cgroup_oom_control_read,
3865 .write_u64 = mem_cgroup_oom_control_write,
3866 .register_event = mem_cgroup_oom_register_event,
3867 .unregister_event = mem_cgroup_oom_unregister_event,
3868 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3872 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3873 static struct cftype memsw_cgroup_files[] = {
3875 .name = "memsw.usage_in_bytes",
3876 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3877 .read_u64 = mem_cgroup_read,
3878 .register_event = mem_cgroup_usage_register_event,
3879 .unregister_event = mem_cgroup_usage_unregister_event,
3882 .name = "memsw.max_usage_in_bytes",
3883 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3884 .trigger = mem_cgroup_reset,
3885 .read_u64 = mem_cgroup_read,
3888 .name = "memsw.limit_in_bytes",
3889 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3890 .write_string = mem_cgroup_write,
3891 .read_u64 = mem_cgroup_read,
3894 .name = "memsw.failcnt",
3895 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3896 .trigger = mem_cgroup_reset,
3897 .read_u64 = mem_cgroup_read,
3901 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3903 if (!do_swap_account)
3905 return cgroup_add_files(cont, ss, memsw_cgroup_files,
3906 ARRAY_SIZE(memsw_cgroup_files));
3909 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3915 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3917 struct mem_cgroup_per_node *pn;
3918 struct mem_cgroup_per_zone *mz;
3920 int zone, tmp = node;
3922 * This routine is called against possible nodes.
3923 * But it's BUG to call kmalloc() against offline node.
3925 * TODO: this routine can waste much memory for nodes which will
3926 * never be onlined. It's better to use memory hotplug callback
3929 if (!node_state(node, N_NORMAL_MEMORY))
3931 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3935 mem->info.nodeinfo[node] = pn;
3936 memset(pn, 0, sizeof(*pn));
3938 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3939 mz = &pn->zoneinfo[zone];
3941 INIT_LIST_HEAD(&mz->lists[l]);
3942 mz->usage_in_excess = 0;
3943 mz->on_tree = false;
3949 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3951 kfree(mem->info.nodeinfo[node]);
3954 static struct mem_cgroup *mem_cgroup_alloc(void)
3956 struct mem_cgroup *mem;
3957 int size = sizeof(struct mem_cgroup);
3959 /* Can be very big if MAX_NUMNODES is very big */
3960 if (size < PAGE_SIZE)
3961 mem = kmalloc(size, GFP_KERNEL);
3963 mem = vmalloc(size);
3968 memset(mem, 0, size);
3969 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3971 if (size < PAGE_SIZE)
3981 * At destroying mem_cgroup, references from swap_cgroup can remain.
3982 * (scanning all at force_empty is too costly...)
3984 * Instead of clearing all references at force_empty, we remember
3985 * the number of reference from swap_cgroup and free mem_cgroup when
3986 * it goes down to 0.
3988 * Removal of cgroup itself succeeds regardless of refs from swap.
3991 static void __mem_cgroup_free(struct mem_cgroup *mem)
3995 mem_cgroup_remove_from_trees(mem);
3996 free_css_id(&mem_cgroup_subsys, &mem->css);
3998 for_each_node_state(node, N_POSSIBLE)
3999 free_mem_cgroup_per_zone_info(mem, node);
4001 free_percpu(mem->stat);
4002 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4008 static void mem_cgroup_get(struct mem_cgroup *mem)
4010 atomic_inc(&mem->refcnt);
4013 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4015 if (atomic_sub_and_test(count, &mem->refcnt)) {
4016 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4017 __mem_cgroup_free(mem);
4019 mem_cgroup_put(parent);
4023 static void mem_cgroup_put(struct mem_cgroup *mem)
4025 __mem_cgroup_put(mem, 1);
4029 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4031 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4033 if (!mem->res.parent)
4035 return mem_cgroup_from_res_counter(mem->res.parent, res);
4038 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4039 static void __init enable_swap_cgroup(void)
4041 if (!mem_cgroup_disabled() && really_do_swap_account)
4042 do_swap_account = 1;
4045 static void __init enable_swap_cgroup(void)
4050 static int mem_cgroup_soft_limit_tree_init(void)
4052 struct mem_cgroup_tree_per_node *rtpn;
4053 struct mem_cgroup_tree_per_zone *rtpz;
4054 int tmp, node, zone;
4056 for_each_node_state(node, N_POSSIBLE) {
4058 if (!node_state(node, N_NORMAL_MEMORY))
4060 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4064 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4066 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4067 rtpz = &rtpn->rb_tree_per_zone[zone];
4068 rtpz->rb_root = RB_ROOT;
4069 spin_lock_init(&rtpz->lock);
4075 static struct cgroup_subsys_state * __ref
4076 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4078 struct mem_cgroup *mem, *parent;
4079 long error = -ENOMEM;
4082 mem = mem_cgroup_alloc();
4084 return ERR_PTR(error);
4086 for_each_node_state(node, N_POSSIBLE)
4087 if (alloc_mem_cgroup_per_zone_info(mem, node))
4091 if (cont->parent == NULL) {
4093 enable_swap_cgroup();
4095 root_mem_cgroup = mem;
4096 if (mem_cgroup_soft_limit_tree_init())
4098 for_each_possible_cpu(cpu) {
4099 struct memcg_stock_pcp *stock =
4100 &per_cpu(memcg_stock, cpu);
4101 INIT_WORK(&stock->work, drain_local_stock);
4103 hotcpu_notifier(memcg_stock_cpu_callback, 0);
4105 parent = mem_cgroup_from_cont(cont->parent);
4106 mem->use_hierarchy = parent->use_hierarchy;
4107 mem->oom_kill_disable = parent->oom_kill_disable;
4110 if (parent && parent->use_hierarchy) {
4111 res_counter_init(&mem->res, &parent->res);
4112 res_counter_init(&mem->memsw, &parent->memsw);
4114 * We increment refcnt of the parent to ensure that we can
4115 * safely access it on res_counter_charge/uncharge.
4116 * This refcnt will be decremented when freeing this
4117 * mem_cgroup(see mem_cgroup_put).
4119 mem_cgroup_get(parent);
4121 res_counter_init(&mem->res, NULL);
4122 res_counter_init(&mem->memsw, NULL);
4124 mem->last_scanned_child = 0;
4125 spin_lock_init(&mem->reclaim_param_lock);
4126 INIT_LIST_HEAD(&mem->oom_notify);
4129 mem->swappiness = get_swappiness(parent);
4130 atomic_set(&mem->refcnt, 1);
4131 mem->move_charge_at_immigrate = 0;
4132 mutex_init(&mem->thresholds_lock);
4135 __mem_cgroup_free(mem);
4136 root_mem_cgroup = NULL;
4137 return ERR_PTR(error);
4140 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4141 struct cgroup *cont)
4143 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4145 return mem_cgroup_force_empty(mem, false);
4148 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4149 struct cgroup *cont)
4151 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4153 mem_cgroup_put(mem);
4156 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4157 struct cgroup *cont)
4161 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4162 ARRAY_SIZE(mem_cgroup_files));
4165 ret = register_memsw_files(cont, ss);
4170 /* Handlers for move charge at task migration. */
4171 #define PRECHARGE_COUNT_AT_ONCE 256
4172 static int mem_cgroup_do_precharge(unsigned long count)
4175 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4176 struct mem_cgroup *mem = mc.to;
4178 if (mem_cgroup_is_root(mem)) {
4179 mc.precharge += count;
4180 /* we don't need css_get for root */
4183 /* try to charge at once */
4185 struct res_counter *dummy;
4187 * "mem" cannot be under rmdir() because we've already checked
4188 * by cgroup_lock_live_cgroup() that it is not removed and we
4189 * are still under the same cgroup_mutex. So we can postpone
4192 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4194 if (do_swap_account && res_counter_charge(&mem->memsw,
4195 PAGE_SIZE * count, &dummy)) {
4196 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4199 mc.precharge += count;
4200 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
4201 WARN_ON_ONCE(count > INT_MAX);
4202 __css_get(&mem->css, (int)count);
4206 /* fall back to one by one charge */
4208 if (signal_pending(current)) {
4212 if (!batch_count--) {
4213 batch_count = PRECHARGE_COUNT_AT_ONCE;
4216 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4218 /* mem_cgroup_clear_mc() will do uncharge later */
4226 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4227 * @vma: the vma the pte to be checked belongs
4228 * @addr: the address corresponding to the pte to be checked
4229 * @ptent: the pte to be checked
4230 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4233 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4234 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4235 * move charge. if @target is not NULL, the page is stored in target->page
4236 * with extra refcnt got(Callers should handle it).
4237 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4238 * target for charge migration. if @target is not NULL, the entry is stored
4241 * Called with pte lock held.
4248 enum mc_target_type {
4249 MC_TARGET_NONE, /* not used */
4254 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4255 unsigned long addr, pte_t ptent)
4257 struct page *page = vm_normal_page(vma, addr, ptent);
4259 if (!page || !page_mapped(page))
4261 if (PageAnon(page)) {
4262 /* we don't move shared anon */
4263 if (!move_anon() || page_mapcount(page) > 2)
4265 } else if (!move_file())
4266 /* we ignore mapcount for file pages */
4268 if (!get_page_unless_zero(page))
4274 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4275 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4278 struct page *page = NULL;
4279 swp_entry_t ent = pte_to_swp_entry(ptent);
4281 if (!move_anon() || non_swap_entry(ent))
4283 usage_count = mem_cgroup_count_swap_user(ent, &page);
4284 if (usage_count > 1) { /* we don't move shared anon */
4289 if (do_swap_account)
4290 entry->val = ent.val;
4295 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4296 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4298 struct page *page = NULL;
4299 struct inode *inode;
4300 struct address_space *mapping;
4303 if (!vma->vm_file) /* anonymous vma */
4308 inode = vma->vm_file->f_path.dentry->d_inode;
4309 mapping = vma->vm_file->f_mapping;
4310 if (pte_none(ptent))
4311 pgoff = linear_page_index(vma, addr);
4312 else /* pte_file(ptent) is true */
4313 pgoff = pte_to_pgoff(ptent);
4315 /* page is moved even if it's not RSS of this task(page-faulted). */
4316 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4317 page = find_get_page(mapping, pgoff);
4318 } else { /* shmem/tmpfs file. we should take account of swap too. */
4320 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4321 if (do_swap_account)
4322 entry->val = ent.val;
4328 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4329 unsigned long addr, pte_t ptent, union mc_target *target)
4331 struct page *page = NULL;
4332 struct page_cgroup *pc;
4334 swp_entry_t ent = { .val = 0 };
4336 if (pte_present(ptent))
4337 page = mc_handle_present_pte(vma, addr, ptent);
4338 else if (is_swap_pte(ptent))
4339 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4340 else if (pte_none(ptent) || pte_file(ptent))
4341 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4343 if (!page && !ent.val)
4346 pc = lookup_page_cgroup(page);
4348 * Do only loose check w/o page_cgroup lock.
4349 * mem_cgroup_move_account() checks the pc is valid or not under
4352 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4353 ret = MC_TARGET_PAGE;
4355 target->page = page;
4357 if (!ret || !target)
4360 /* There is a swap entry and a page doesn't exist or isn't charged */
4361 if (ent.val && !ret &&
4362 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4363 ret = MC_TARGET_SWAP;
4370 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4371 unsigned long addr, unsigned long end,
4372 struct mm_walk *walk)
4374 struct vm_area_struct *vma = walk->private;
4378 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4379 for (; addr != end; pte++, addr += PAGE_SIZE)
4380 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4381 mc.precharge++; /* increment precharge temporarily */
4382 pte_unmap_unlock(pte - 1, ptl);
4388 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4390 unsigned long precharge;
4391 struct vm_area_struct *vma;
4393 down_read(&mm->mmap_sem);
4394 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4395 struct mm_walk mem_cgroup_count_precharge_walk = {
4396 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4400 if (is_vm_hugetlb_page(vma))
4402 walk_page_range(vma->vm_start, vma->vm_end,
4403 &mem_cgroup_count_precharge_walk);
4405 up_read(&mm->mmap_sem);
4407 precharge = mc.precharge;
4413 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4415 return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4418 static void mem_cgroup_clear_mc(void)
4420 /* we must uncharge all the leftover precharges from mc.to */
4422 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4424 memcg_oom_recover(mc.to);
4427 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4428 * we must uncharge here.
4430 if (mc.moved_charge) {
4431 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4432 mc.moved_charge = 0;
4433 memcg_oom_recover(mc.from);
4435 /* we must fixup refcnts and charges */
4436 if (mc.moved_swap) {
4437 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4438 /* uncharge swap account from the old cgroup */
4439 if (!mem_cgroup_is_root(mc.from))
4440 res_counter_uncharge(&mc.from->memsw,
4441 PAGE_SIZE * mc.moved_swap);
4442 __mem_cgroup_put(mc.from, mc.moved_swap);
4444 if (!mem_cgroup_is_root(mc.to)) {
4446 * we charged both to->res and to->memsw, so we should
4449 res_counter_uncharge(&mc.to->res,
4450 PAGE_SIZE * mc.moved_swap);
4451 VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4452 __css_put(&mc.to->css, mc.moved_swap);
4454 /* we've already done mem_cgroup_get(mc.to) */
4460 mc.moving_task = NULL;
4461 wake_up_all(&mc.waitq);
4464 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4465 struct cgroup *cgroup,
4466 struct task_struct *p,
4470 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4472 if (mem->move_charge_at_immigrate) {
4473 struct mm_struct *mm;
4474 struct mem_cgroup *from = mem_cgroup_from_task(p);
4476 VM_BUG_ON(from == mem);
4478 mm = get_task_mm(p);
4481 /* We move charges only when we move a owner of the mm */
4482 if (mm->owner == p) {
4485 VM_BUG_ON(mc.precharge);
4486 VM_BUG_ON(mc.moved_charge);
4487 VM_BUG_ON(mc.moved_swap);
4488 VM_BUG_ON(mc.moving_task);
4492 mc.moved_charge = 0;
4494 mc.moving_task = current;
4496 ret = mem_cgroup_precharge_mc(mm);
4498 mem_cgroup_clear_mc();
4505 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4506 struct cgroup *cgroup,
4507 struct task_struct *p,
4510 mem_cgroup_clear_mc();
4513 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4514 unsigned long addr, unsigned long end,
4515 struct mm_walk *walk)
4518 struct vm_area_struct *vma = walk->private;
4523 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4524 for (; addr != end; addr += PAGE_SIZE) {
4525 pte_t ptent = *(pte++);
4526 union mc_target target;
4529 struct page_cgroup *pc;
4535 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4537 case MC_TARGET_PAGE:
4539 if (isolate_lru_page(page))
4541 pc = lookup_page_cgroup(page);
4542 if (!mem_cgroup_move_account(pc,
4543 mc.from, mc.to, false)) {
4545 /* we uncharge from mc.from later. */
4548 putback_lru_page(page);
4549 put: /* is_target_pte_for_mc() gets the page */
4552 case MC_TARGET_SWAP:
4554 if (!mem_cgroup_move_swap_account(ent,
4555 mc.from, mc.to, false)) {
4557 /* we fixup refcnts and charges later. */
4565 pte_unmap_unlock(pte - 1, ptl);
4570 * We have consumed all precharges we got in can_attach().
4571 * We try charge one by one, but don't do any additional
4572 * charges to mc.to if we have failed in charge once in attach()
4575 ret = mem_cgroup_do_precharge(1);
4583 static void mem_cgroup_move_charge(struct mm_struct *mm)
4585 struct vm_area_struct *vma;
4587 lru_add_drain_all();
4588 down_read(&mm->mmap_sem);
4589 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4591 struct mm_walk mem_cgroup_move_charge_walk = {
4592 .pmd_entry = mem_cgroup_move_charge_pte_range,
4596 if (is_vm_hugetlb_page(vma))
4598 ret = walk_page_range(vma->vm_start, vma->vm_end,
4599 &mem_cgroup_move_charge_walk);
4602 * means we have consumed all precharges and failed in
4603 * doing additional charge. Just abandon here.
4607 up_read(&mm->mmap_sem);
4610 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4611 struct cgroup *cont,
4612 struct cgroup *old_cont,
4613 struct task_struct *p,
4616 struct mm_struct *mm;
4619 /* no need to move charge */
4622 mm = get_task_mm(p);
4624 mem_cgroup_move_charge(mm);
4627 mem_cgroup_clear_mc();
4629 #else /* !CONFIG_MMU */
4630 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4631 struct cgroup *cgroup,
4632 struct task_struct *p,
4637 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4638 struct cgroup *cgroup,
4639 struct task_struct *p,
4643 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4644 struct cgroup *cont,
4645 struct cgroup *old_cont,
4646 struct task_struct *p,
4652 struct cgroup_subsys mem_cgroup_subsys = {
4654 .subsys_id = mem_cgroup_subsys_id,
4655 .create = mem_cgroup_create,
4656 .pre_destroy = mem_cgroup_pre_destroy,
4657 .destroy = mem_cgroup_destroy,
4658 .populate = mem_cgroup_populate,
4659 .can_attach = mem_cgroup_can_attach,
4660 .cancel_attach = mem_cgroup_cancel_attach,
4661 .attach = mem_cgroup_move_task,
4666 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4668 static int __init disable_swap_account(char *s)
4670 really_do_swap_account = 0;
4673 __setup("noswapaccount", disable_swap_account);